Research Of Fine-grained Image Recognition Based On Attention Mechanism And Adversarial Loss

Fine-grained image recognition focuses on dealing with the image belonging to the sub-categories of one meta-category.With the technology of image recognition applies in various fields,people’s demand for fine-grained object category has increased.Therefore,the work of fine-grained image recognition has gradually become a hot topic of research.Aiming at the problem of fine-grained image recognition with the small inter-class variations and the large intra-class variations,recent works mainly tackle this problem by focusing on how to learn the most discriminative feature of the image.Therefore,the research of fine-grained image recognition can be generally classified into three dimensions: fine-grained feature learning,discriminative part localization and the strategy of data augmentation.This paper analyzes and conducts the research of the discriminative part localization and the strategy of data augmentation in detail.Fine-grained image recognition based on channel attention mechanism and region augmentation is proposed.Firstly,the method uses channel attention mechanism to obtain the key region of the original image.Secondly,the method can force the network to focus on the detailed region of fine-grained image by conducting the region augmentation.Finally,the method constructs concatenate model in the target region and local regions on weakly-supervised datasets.This method combines the attention mechanism and the strategy of data augmentation.Therefore,the method can not only minimize the influence of background noise but also pay attention to the detailed region of the finegrained image.The experimental results on multiple benchmark datasets show that this method is effective to improve the accuracy of fine-grained image recognition.Fine-grained image recognition method based on multi-grained adversarial loss is proposed.Firstly,the method utilizes the strategy of data augmentation to build the multi-grinded image in original one to highlight the detailed information of the fine-grinded image.Then,the network fuses the feature of the multi-grinded image by the strategy of progressive training.Finally,the method designs the delicate-designed multi-grinded adversarial loss function,which can erase the influence caused by the noise which comes from the destruction of space information.Therefore,this method can effectively learn the multi-scale features from local information.The experiment results show that our method effective to learn the discriminative features to improve the accuracy of recognition.